experiments/status-aware/logs/ablations/humanized_structdense.log

============================================================
  StructGNN (act=64d hash, pos=0d) [NO GNN] [adapted]: kuhperdata-exp
============================================================
Loading pre-computed data...
/workspace/ta-statute-law-retrieval/src/evaluate_paragnn.py:140: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  bm25_train_scores = torch.load(f"{output_dir}/bm25_train_scores.pt")
/workspace/ta-statute-law-retrieval/src/evaluate_paragnn.py:141: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  bm25_val_scores = torch.load(f"{output_dir}/bm25_val_scores.pt")
/workspace/ta-statute-law-retrieval/src/evaluate_paragnn.py:142: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  bm25_test_scores = torch.load(f"{output_dir}/bm25_test_scores.pt")
Loading paragraph store...
/workspace/ta-statute-law-retrieval/src/paragnn/graph_builder.py:40: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  self.rr_const_emb = torch.load(self.emb_dir / "EMBD_CONST.pt")  # (13, 1024)
Computing structure features (act_encoder=hash, act_feat=64d, pos=0d)...
  2127 corpus docs, feature dim=64
Creating training dataset...
Creating ParaGNN training dataset...
  Created 2534 training examples from 1035 queries
Building val graph...
/workspace/ta-statute-law-retrieval/src/paragnn/graph_builder.py:81: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  self._emb_cache[qid] = torch.load(path, map_location="cpu")
/workspace/ta-statute-law-retrieval/src/paragnn/graph_builder.py:88: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  self._emb_cache[key] = torch.load(path, map_location="cpu")
Building test graph...
Training StructGNN for epochs 1-100...
  Early stopping on VAL set (148 queries)
Pre-warming corpus embedding cache (2127 docs)...
  Cache ready.

Epoch 1:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 1:  10%|█         | 1/10 [00:02<00:22,  2.48s/it]
Epoch 1:  30%|███       | 3/10 [00:03<00:06,  1.06it/s]
Epoch 1:  50%|█████     | 5/10 [00:04<00:04,  1.24it/s]
Epoch 1:  60%|██████    | 6/10 [00:04<00:02,  1.61it/s]
Epoch 1:  70%|███████   | 7/10 [00:05<00:02,  1.34it/s]
Epoch 1:  80%|████████  | 8/10 [00:06<00:01,  1.29it/s]
Epoch 1:  90%|█████████ | 9/10 [00:06<00:00,  1.61it/s]
Epoch 1: 100%|██████████| 10/10 [00:07<00:00,  1.69it/s]
                                                        
  Epoch 1: loss=2.5772 val_MRR=0.0695 val_R@10=0.1103 val_Hit=18.9% alpha=0.8 (learned=0.582)
  → New best val Recall@10=0.1103, saved model

Epoch 2:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 2:  10%|█         | 1/10 [00:01<00:17,  1.93s/it]
Epoch 2:  30%|███       | 3/10 [00:03<00:06,  1.00it/s]
Epoch 2:  40%|████      | 4/10 [00:03<00:04,  1.38it/s]
Epoch 2:  50%|█████     | 5/10 [00:04<00:04,  1.03it/s]
Epoch 2:  60%|██████    | 6/10 [00:05<00:02,  1.37it/s]
Epoch 2:  70%|███████   | 7/10 [00:06<00:02,  1.15it/s]
Epoch 2:  80%|████████  | 8/10 [00:06<00:01,  1.57it/s]
Epoch 2:  90%|█████████ | 9/10 [00:07<00:00,  1.38it/s]
Epoch 2: 100%|██████████| 10/10 [00:07<00:00,  1.83it/s]
                                                        
  Epoch 2: loss=2.2010 val_MRR=0.0695 val_R@10=0.1126 val_Hit=18.9% alpha=0.8 (learned=0.812)
  → New best val Recall@10=0.1126, saved model

Epoch 3:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 3:  10%|█         | 1/10 [00:02<00:21,  2.38s/it]
Epoch 3:  30%|███       | 3/10 [00:03<00:07,  1.09s/it]
Epoch 3:  50%|█████     | 5/10 [00:05<00:04,  1.03it/s]
Epoch 3:  60%|██████    | 6/10 [00:05<00:02,  1.34it/s]
Epoch 3:  70%|███████   | 7/10 [00:07<00:03,  1.02s/it]
Epoch 3:  90%|█████████ | 9/10 [00:08<00:00,  1.23it/s]
                                                       
  Epoch 3: loss=1.7966 val_MRR=0.1165 val_R@10=0.1950 val_Hit=34.5% alpha=0.9 (learned=0.934)
  → New best val Recall@10=0.1950, saved model

Epoch 4:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 4:  10%|█         | 1/10 [00:01<00:17,  1.92s/it]
Epoch 4:  30%|███       | 3/10 [00:03<00:07,  1.11s/it]
Epoch 4:  50%|█████     | 5/10 [00:05<00:05,  1.02s/it]
Epoch 4:  60%|██████    | 6/10 [00:05<00:03,  1.29it/s]
Epoch 4:  70%|███████   | 7/10 [00:06<00:02,  1.15it/s]
Epoch 4:  90%|█████████ | 9/10 [00:07<00:00,  1.46it/s]
                                                       
  Epoch 4: loss=1.4887 val_MRR=0.2892 val_R@10=0.3501 val_Hit=53.4% alpha=0.9 (learned=0.969)
  → New best val Recall@10=0.3501, saved model

Epoch 5:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 5:  10%|█         | 1/10 [00:02<00:22,  2.52s/it]
Epoch 5:  30%|███       | 3/10 [00:04<00:08,  1.23s/it]
Epoch 5:  50%|█████     | 5/10 [00:05<00:04,  1.07it/s]
Epoch 5:  70%|███████   | 7/10 [00:06<00:02,  1.16it/s]
Epoch 5:  90%|█████████ | 9/10 [00:07<00:00,  1.45it/s]
                                                       
  Epoch 5: loss=1.3158 val_MRR=0.3553 val_R@10=0.3630 val_Hit=54.7% alpha=0.8 (learned=0.980)
  → New best val Recall@10=0.3630, saved model

Epoch 6:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 6:  10%|█         | 1/10 [00:02<00:23,  2.66s/it]
Epoch 6:  30%|███       | 3/10 [00:04<00:08,  1.25s/it]
Epoch 6:  40%|████      | 4/10 [00:04<00:05,  1.14it/s]
Epoch 6:  50%|█████     | 5/10 [00:05<00:04,  1.02it/s]
Epoch 6:  70%|███████   | 7/10 [00:07<00:02,  1.13it/s]
Epoch 6:  90%|█████████ | 9/10 [00:08<00:00,  1.35it/s]
                                                       
  Epoch 6: loss=1.2068 val_MRR=0.3641 val_R@10=0.4127 val_Hit=58.8% alpha=0.8 (learned=0.984)
  → New best val Recall@10=0.4127, saved model

Epoch 7:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 7:  10%|█         | 1/10 [00:06<01:01,  6.88s/it]
Epoch 7:  20%|██        | 2/10 [00:08<00:30,  3.78s/it]
Epoch 7:  30%|███       | 3/10 [00:10<00:19,  2.84s/it]
Epoch 7:  40%|████      | 4/10 [00:10<00:10,  1.78s/it]
Epoch 7:  50%|█████     | 5/10 [00:11<00:07,  1.53s/it]
Epoch 7:  60%|██████    | 6/10 [00:11<00:04,  1.12s/it]
Epoch 7:  70%|███████   | 7/10 [00:13<00:03,  1.17s/it]
Epoch 7:  80%|████████  | 8/10 [00:13<00:01,  1.02it/s]
Epoch 7:  90%|█████████ | 9/10 [00:14<00:00,  1.20it/s]
Epoch 7: 100%|██████████| 10/10 [00:14<00:00,  1.46it/s]
                                                        
  Epoch 7: loss=1.1577 val_MRR=0.2876 val_R@10=0.4397 val_Hit=59.5% alpha=0.9 (learned=0.988)
  → New best val Recall@10=0.4397, saved model

Epoch 8:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 8:  10%|█         | 1/10 [00:02<00:19,  2.14s/it]
Epoch 8:  20%|██        | 2/10 [00:02<00:08,  1.11s/it]
Epoch 8:  30%|███       | 3/10 [00:03<00:07,  1.11s/it]
Epoch 8:  40%|████      | 4/10 [00:03<00:04,  1.36it/s]
Epoch 8:  50%|█████     | 5/10 [00:05<00:04,  1.01it/s]
Epoch 8:  60%|██████    | 6/10 [00:05<00:02,  1.41it/s]
Epoch 8:  70%|███████   | 7/10 [00:06<00:02,  1.19it/s]
Epoch 8:  90%|█████████ | 9/10 [00:07<00:00,  1.29it/s]
                                                       
  Epoch 8: loss=1.0931 val_MRR=0.4104 val_R@10=0.4594 val_Hit=62.8% alpha=0.8 (learned=0.992)
  → New best val Recall@10=0.4594, saved model

Epoch 9:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 9:  10%|█         | 1/10 [00:02<00:18,  2.01s/it]
Epoch 9:  30%|███       | 3/10 [00:03<00:08,  1.24s/it]
Epoch 9:  50%|█████     | 5/10 [00:05<00:04,  1.08it/s]
Epoch 9:  70%|███████   | 7/10 [00:06<00:02,  1.11it/s]
Epoch 9:  90%|█████████ | 9/10 [00:08<00:00,  1.29it/s]
                                                       
  Epoch 9: loss=1.0613 val_MRR=0.3263 val_R@10=0.5447 val_Hit=67.6% alpha=0.9 (learned=0.992)
  → New best val Recall@10=0.5447, saved model

Epoch 10:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 10:  10%|█         | 1/10 [00:02<00:20,  2.32s/it]
Epoch 10:  20%|██        | 2/10 [00:02<00:08,  1.06s/it]
Epoch 10:  30%|███       | 3/10 [00:03<00:08,  1.25s/it]
Epoch 10:  40%|████      | 4/10 [00:04<00:04,  1.20it/s]
Epoch 10:  50%|█████     | 5/10 [00:05<00:05,  1.05s/it]
Epoch 10:  60%|██████    | 6/10 [00:05<00:03,  1.30it/s]
Epoch 10:  70%|███████   | 7/10 [00:07<00:02,  1.11it/s]
Epoch 10:  80%|████████  | 8/10 [00:07<00:01,  1.49it/s]
Epoch 10:  90%|█████████ | 9/10 [00:07<00:00,  1.46it/s]
                                                        
  Epoch 10: loss=1.0171 val_MRR=0.4488 val_R@10=0.6797 val_Hit=80.4% alpha=0.9 (learned=0.992)
  → New best val Recall@10=0.6797, saved model

Epoch 11:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 11:  10%|█         | 1/10 [00:02<00:25,  2.79s/it]
Epoch 11:  30%|███       | 3/10 [00:04<00:09,  1.41s/it]
Epoch 11:  50%|█████     | 5/10 [00:06<00:05,  1.04s/it]
Epoch 11:  70%|███████   | 7/10 [00:07<00:02,  1.08it/s]
Epoch 11:  90%|█████████ | 9/10 [00:08<00:00,  1.36it/s]
                                                        
  Epoch 11: loss=0.9840 val_MRR=0.5126 val_R@10=0.7752 val_Hit=89.2% alpha=0.9 (learned=0.992)
  → New best val Recall@10=0.7752, saved model

Epoch 12:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 12:  10%|█         | 1/10 [00:02<00:23,  2.56s/it]
Epoch 12:  30%|███       | 3/10 [00:04<00:08,  1.28s/it]
Epoch 12:  50%|█████     | 5/10 [00:05<00:05,  1.06s/it]
Epoch 12:  70%|███████   | 7/10 [00:07<00:02,  1.13it/s]
Epoch 12:  90%|█████████ | 9/10 [00:08<00:00,  1.40it/s]
                                                        
  Epoch 12: loss=0.9502 val_MRR=0.5194 val_R@10=0.8211 val_Hit=91.9% alpha=0.9 (learned=0.992)
  → New best val Recall@10=0.8211, saved model

Epoch 13:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 13:  10%|█         | 1/10 [00:02<00:21,  2.34s/it]
Epoch 13:  30%|███       | 3/10 [00:03<00:07,  1.08s/it]
Epoch 13:  50%|█████     | 5/10 [00:04<00:04,  1.16it/s]
Epoch 13:  70%|███████   | 7/10 [00:06<00:02,  1.22it/s]
Epoch 13:  90%|█████████ | 9/10 [00:07<00:00,  1.40it/s]
                                                        
  Epoch 13: loss=0.8989 val_MRR=0.5629 val_R@10=0.8413 val_Hit=93.2% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.8413, saved model

Epoch 14:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 14:  10%|█         | 1/10 [00:02<00:25,  2.78s/it]
Epoch 14:  30%|███       | 3/10 [00:04<00:08,  1.20s/it]
Epoch 14:  50%|█████     | 5/10 [00:05<00:05,  1.06s/it]
Epoch 14:  70%|███████   | 7/10 [00:07<00:02,  1.14it/s]
Epoch 14:  90%|█████████ | 9/10 [00:08<00:00,  1.43it/s]
                                                        
  Epoch 14: loss=0.8549 val_MRR=0.6084 val_R@10=0.8616 val_Hit=94.6% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.8616, saved model

Epoch 15:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 15:  10%|█         | 1/10 [00:02<00:22,  2.45s/it]
Epoch 15:  30%|███       | 3/10 [00:03<00:08,  1.21s/it]
Epoch 15:  50%|█████     | 5/10 [00:05<00:04,  1.10it/s]
Epoch 15:  70%|███████   | 7/10 [00:06<00:02,  1.26it/s]
Epoch 15:  90%|█████████ | 9/10 [00:07<00:00,  1.53it/s]
Epoch 15: 100%|██████████| 10/10 [00:07<00:00,  1.73it/s]
                                                         
  Epoch 15: loss=0.8485 val_MRR=0.6304 val_R@10=0.8543 val_Hit=93.9% alpha=0.9 (learned=0.996)

Epoch 16:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 16:  10%|█         | 1/10 [00:02<00:23,  2.59s/it]
Epoch 16:  20%|██        | 2/10 [00:02<00:09,  1.14s/it]
Epoch 16:  30%|███       | 3/10 [00:03<00:08,  1.18s/it]
Epoch 16:  40%|████      | 4/10 [00:04<00:04,  1.31it/s]
Epoch 16:  50%|█████     | 5/10 [00:05<00:04,  1.10it/s]
Epoch 16:  60%|██████    | 6/10 [00:05<00:02,  1.54it/s]
Epoch 16:  70%|███████   | 7/10 [00:06<00:02,  1.11it/s]
Epoch 16:  90%|█████████ | 9/10 [00:07<00:00,  1.40it/s]
                                                        
  Epoch 16: loss=0.8084 val_MRR=0.6393 val_R@10=0.8684 val_Hit=95.3% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.8684, saved model

Epoch 17:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 17:  10%|█         | 1/10 [00:02<00:21,  2.36s/it]
Epoch 17:  20%|██        | 2/10 [00:02<00:08,  1.04s/it]
Epoch 17:  30%|███       | 3/10 [00:03<00:07,  1.10s/it]
Epoch 17:  50%|█████     | 5/10 [00:05<00:04,  1.09it/s]
Epoch 17:  70%|███████   | 7/10 [00:06<00:02,  1.22it/s]
Epoch 17:  90%|█████████ | 9/10 [00:07<00:00,  1.51it/s]
                                                        
  Epoch 17: loss=0.7891 val_MRR=0.6606 val_R@10=0.8650 val_Hit=94.6% alpha=0.9 (learned=0.996)

Epoch 18:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 18:  10%|█         | 1/10 [00:02<00:22,  2.46s/it]
Epoch 18:  30%|███       | 3/10 [00:04<00:09,  1.32s/it]
Epoch 18:  50%|█████     | 5/10 [00:05<00:04,  1.04it/s]
Epoch 18:  70%|███████   | 7/10 [00:06<00:02,  1.23it/s]
Epoch 18:  90%|█████████ | 9/10 [00:07<00:00,  1.53it/s]
                                                        
  Epoch 18: loss=0.7799 val_MRR=0.6699 val_R@10=0.8762 val_Hit=95.3% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.8762, saved model

Epoch 19:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 19:  10%|█         | 1/10 [00:02<00:26,  2.95s/it]
Epoch 19:  30%|███       | 3/10 [00:04<00:08,  1.24s/it]
Epoch 19:  50%|█████     | 5/10 [00:05<00:04,  1.09it/s]
Epoch 19:  60%|██████    | 6/10 [00:05<00:02,  1.39it/s]
Epoch 19:  70%|███████   | 7/10 [00:06<00:02,  1.18it/s]
Epoch 19:  90%|█████████ | 9/10 [00:08<00:00,  1.34it/s]
                                                        
  Epoch 19: loss=0.7536 val_MRR=0.6987 val_R@10=0.8920 val_Hit=96.6% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.8920, saved model

Epoch 20:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 20:  10%|█         | 1/10 [00:02<00:24,  2.77s/it]
Epoch 20:  30%|███       | 3/10 [00:03<00:08,  1.15s/it]
Epoch 20:  50%|█████     | 5/10 [00:05<00:04,  1.13it/s]
Epoch 20:  70%|███████   | 7/10 [00:07<00:02,  1.10it/s]
Epoch 20:  90%|█████████ | 9/10 [00:08<00:00,  1.35it/s]
                                                        
  Epoch 20: loss=0.7386 val_MRR=0.6822 val_R@10=0.8898 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 21:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 21:  10%|█         | 1/10 [00:02<00:22,  2.45s/it]
Epoch 21:  30%|███       | 3/10 [00:04<00:09,  1.30s/it]
Epoch 21:  50%|█████     | 5/10 [00:05<00:05,  1.01s/it]
Epoch 21:  70%|███████   | 7/10 [00:07<00:02,  1.15it/s]
Epoch 21:  90%|█████████ | 9/10 [00:07<00:00,  1.45it/s]
                                                        
  Epoch 21: loss=0.6928 val_MRR=0.7075 val_R@10=0.8929 val_Hit=96.6% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.8929, saved model

Epoch 22:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 22:  10%|█         | 1/10 [00:03<00:27,  3.11s/it]
Epoch 22:  20%|██        | 2/10 [00:04<00:17,  2.15s/it]
Epoch 22:  30%|███       | 3/10 [00:10<00:27,  3.91s/it]
Epoch 22:  40%|████      | 4/10 [00:14<00:22,  3.75s/it]
Epoch 22:  50%|█████     | 5/10 [00:20<00:24,  4.83s/it]
Epoch 22:  60%|██████    | 6/10 [00:20<00:12,  3.23s/it]
Epoch 22:  70%|███████   | 7/10 [00:22<00:07,  2.53s/it]
Epoch 22:  90%|█████████ | 9/10 [00:22<00:01,  1.50s/it]
                                                        
  Epoch 22: loss=0.6842 val_MRR=0.7020 val_R@10=0.9021 val_Hit=97.3% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.9021, saved model

Epoch 23:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 23:  10%|█         | 1/10 [00:01<00:17,  1.99s/it]
Epoch 23:  30%|███       | 3/10 [00:04<00:09,  1.41s/it]
Epoch 23:  40%|████      | 4/10 [00:04<00:05,  1.03it/s]
Epoch 23:  50%|█████     | 5/10 [00:05<00:05,  1.12s/it]
Epoch 23:  70%|███████   | 7/10 [00:07<00:02,  1.17it/s]
Epoch 23:  90%|█████████ | 9/10 [00:07<00:00,  1.48it/s]
                                                        
  Epoch 23: loss=0.6618 val_MRR=0.6972 val_R@10=0.8819 val_Hit=95.3% alpha=0.9 (learned=0.996)

Epoch 24:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 24:  10%|█         | 1/10 [00:03<00:31,  3.46s/it]
Epoch 24:  30%|███       | 3/10 [00:04<00:09,  1.35s/it]
Epoch 24:  40%|████      | 4/10 [00:04<00:05,  1.08it/s]
Epoch 24:  50%|█████     | 5/10 [00:06<00:05,  1.10s/it]
Epoch 24:  60%|██████    | 6/10 [00:06<00:03,  1.27it/s]
Epoch 24:  70%|███████   | 7/10 [00:07<00:02,  1.01it/s]
Epoch 24:  90%|█████████ | 9/10 [00:08<00:00,  1.38it/s]
                                                        
  Epoch 24: loss=0.6427 val_MRR=0.7070 val_R@10=0.8943 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 25:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 25:  10%|█         | 1/10 [00:02<00:23,  2.61s/it]
Epoch 25:  30%|███       | 3/10 [00:03<00:07,  1.10s/it]
Epoch 25:  40%|████      | 4/10 [00:04<00:05,  1.18it/s]
Epoch 25:  50%|█████     | 5/10 [00:05<00:04,  1.01it/s]
Epoch 25:  60%|██████    | 6/10 [00:05<00:03,  1.26it/s]
Epoch 25:  70%|███████   | 7/10 [00:06<00:02,  1.08it/s]
Epoch 25:  80%|████████  | 8/10 [00:07<00:01,  1.30it/s]
Epoch 25:  90%|█████████ | 9/10 [00:07<00:00,  1.50it/s]
Epoch 25: 100%|██████████| 10/10 [00:08<00:00,  1.82it/s]
                                                         
  Epoch 25: loss=0.6363 val_MRR=0.7195 val_R@10=0.9005 val_Hit=97.3% alpha=0.9 (learned=0.996)

Epoch 26:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 26:  10%|█         | 1/10 [00:01<00:17,  1.99s/it]
Epoch 26:  30%|███       | 3/10 [00:03<00:06,  1.05it/s]
Epoch 26:  50%|█████     | 5/10 [00:04<00:04,  1.19it/s]
Epoch 26:  70%|███████   | 7/10 [00:05<00:02,  1.33it/s]
Epoch 26:  90%|█████████ | 9/10 [00:06<00:00,  1.59it/s]
                                                        
  Epoch 26: loss=0.6049 val_MRR=0.7023 val_R@10=0.8898 val_Hit=95.9% alpha=0.9 (learned=0.996)

Epoch 27:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 27:  10%|█         | 1/10 [00:02<00:18,  2.07s/it]
Epoch 27:  30%|███       | 3/10 [00:03<00:07,  1.06s/it]
Epoch 27:  50%|█████     | 5/10 [00:04<00:04,  1.22it/s]
Epoch 27:  60%|██████    | 6/10 [00:04<00:02,  1.56it/s]
Epoch 27:  70%|███████   | 7/10 [00:05<00:02,  1.33it/s]
Epoch 27:  90%|█████████ | 9/10 [00:06<00:00,  1.67it/s]
                                                        
  Epoch 27: loss=0.6191 val_MRR=0.7133 val_R@10=0.8976 val_Hit=97.3% alpha=0.9 (learned=0.996)

Epoch 28:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 28:  10%|█         | 1/10 [00:02<00:19,  2.17s/it]
Epoch 28:  30%|███       | 3/10 [00:03<00:07,  1.06s/it]
Epoch 28:  50%|█████     | 5/10 [00:04<00:04,  1.18it/s]
Epoch 28:  70%|███████   | 7/10 [00:06<00:02,  1.34it/s]
Epoch 28:  90%|█████████ | 9/10 [00:07<00:00,  1.51it/s]
                                                        
  Epoch 28: loss=0.5941 val_MRR=0.7062 val_R@10=0.8971 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 29:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 29:  10%|█         | 1/10 [00:01<00:16,  1.84s/it]
Epoch 29:  20%|██        | 2/10 [00:02<00:06,  1.16it/s]
Epoch 29:  30%|███       | 3/10 [00:03<00:07,  1.00s/it]
Epoch 29:  40%|████      | 4/10 [00:03<00:03,  1.50it/s]
Epoch 29:  50%|█████     | 5/10 [00:04<00:04,  1.24it/s]
Epoch 29:  60%|██████    | 6/10 [00:04<00:02,  1.52it/s]
Epoch 29:  70%|███████   | 7/10 [00:05<00:02,  1.27it/s]
Epoch 29:  90%|█████████ | 9/10 [00:06<00:00,  1.65it/s]
                                                        
  Epoch 29: loss=0.5852 val_MRR=0.7049 val_R@10=0.8943 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 30:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 30:  10%|█         | 1/10 [00:01<00:16,  1.86s/it]
Epoch 30:  30%|███       | 3/10 [00:03<00:07,  1.08s/it]
Epoch 30:  50%|█████     | 5/10 [00:04<00:04,  1.18it/s]
Epoch 30:  70%|███████   | 7/10 [00:05<00:02,  1.36it/s]
Epoch 30:  90%|█████████ | 9/10 [00:06<00:00,  1.66it/s]
                                                        
  Epoch 30: loss=0.5705 val_MRR=0.7136 val_R@10=0.9038 val_Hit=97.3% alpha=0.9 (learned=0.996)
  → New best val Recall@10=0.9038, saved model

Epoch 31:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 31:  10%|█         | 1/10 [00:02<00:23,  2.65s/it]
Epoch 31:  30%|███       | 3/10 [00:04<00:08,  1.21s/it]
Epoch 31:  50%|█████     | 5/10 [00:05<00:04,  1.10it/s]
Epoch 31:  70%|███████   | 7/10 [00:06<00:02,  1.22it/s]
Epoch 31:  90%|█████████ | 9/10 [00:07<00:00,  1.53it/s]
                                                        
  Epoch 31: loss=0.5456 val_MRR=0.6912 val_R@10=0.8822 val_Hit=95.9% alpha=0.9 (learned=0.996)

Epoch 32:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 32:  10%|█         | 1/10 [00:02<00:26,  2.97s/it]
Epoch 32:  30%|███       | 3/10 [00:04<00:10,  1.46s/it]
Epoch 32:  50%|█████     | 5/10 [00:06<00:05,  1.05s/it]
Epoch 32:  70%|███████   | 7/10 [00:07<00:02,  1.08it/s]
Epoch 32:  90%|█████████ | 9/10 [00:08<00:00,  1.35it/s]
                                                        
  Epoch 32: loss=0.5409 val_MRR=0.6968 val_R@10=0.8971 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 33:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 33:  10%|█         | 1/10 [00:02<00:18,  2.10s/it]
Epoch 33:  30%|███       | 3/10 [00:03<00:07,  1.08s/it]
Epoch 33:  50%|█████     | 5/10 [00:05<00:04,  1.09it/s]
Epoch 33:  70%|███████   | 7/10 [00:06<00:02,  1.28it/s]
Epoch 33:  90%|█████████ | 9/10 [00:07<00:00,  1.47it/s]
                                                        
  Epoch 33: loss=0.5227 val_MRR=0.6984 val_R@10=0.9025 val_Hit=97.3% alpha=0.9 (learned=0.996)

Epoch 34:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 34:  10%|█         | 1/10 [00:01<00:16,  1.84s/it]
Epoch 34:  30%|███       | 3/10 [00:03<00:07,  1.08s/it]
Epoch 34:  50%|█████     | 5/10 [00:04<00:04,  1.18it/s]
Epoch 34:  70%|███████   | 7/10 [00:06<00:02,  1.18it/s]
Epoch 34:  90%|█████████ | 9/10 [00:07<00:00,  1.48it/s]
                                                        
  Epoch 34: loss=0.5069 val_MRR=0.6912 val_R@10=0.8841 val_Hit=95.9% alpha=0.8 (learned=0.996)

Epoch 35:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 35:  10%|█         | 1/10 [00:02<00:18,  2.01s/it]
Epoch 35:  30%|███       | 3/10 [00:03<00:07,  1.10s/it]
Epoch 35:  50%|█████     | 5/10 [00:05<00:04,  1.08it/s]
Epoch 35:  70%|███████   | 7/10 [00:06<00:02,  1.27it/s]
Epoch 35:  90%|█████████ | 9/10 [00:07<00:00,  1.57it/s]
                                                        
  Epoch 35: loss=0.5041 val_MRR=0.6894 val_R@10=0.8935 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 36:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 36:  10%|█         | 1/10 [00:02<00:19,  2.16s/it]
Epoch 36:  20%|██        | 2/10 [00:02<00:07,  1.02it/s]
Epoch 36:  30%|███       | 3/10 [00:03<00:07,  1.14s/it]
Epoch 36:  40%|████      | 4/10 [00:04<00:05,  1.15it/s]
Epoch 36:  50%|█████     | 5/10 [00:05<00:05,  1.01s/it]
Epoch 36:  60%|██████    | 6/10 [00:05<00:02,  1.36it/s]
Epoch 36:  70%|███████   | 7/10 [00:06<00:02,  1.20it/s]
Epoch 36:  80%|████████  | 8/10 [00:06<00:01,  1.49it/s]
Epoch 36:  90%|█████████ | 9/10 [00:07<00:00,  1.49it/s]
                                                        
  Epoch 36: loss=0.4940 val_MRR=0.6933 val_R@10=0.8957 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 37:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 37:  10%|█         | 1/10 [00:01<00:17,  1.96s/it]
Epoch 37:  30%|███       | 3/10 [00:03<00:06,  1.08it/s]
Epoch 37:  40%|████      | 4/10 [00:03<00:03,  1.53it/s]
Epoch 37:  50%|█████     | 5/10 [00:04<00:04,  1.21it/s]
Epoch 37:  60%|██████    | 6/10 [00:04<00:02,  1.54it/s]
Epoch 37:  70%|███████   | 7/10 [00:05<00:02,  1.26it/s]
Epoch 37:  90%|█████████ | 9/10 [00:06<00:00,  1.52it/s]
                                                        
  Epoch 37: loss=0.4824 val_MRR=0.6908 val_R@10=0.8881 val_Hit=95.9% alpha=0.9 (learned=0.996)

Epoch 38:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 38:  10%|█         | 1/10 [00:01<00:16,  1.80s/it]
Epoch 38:  30%|███       | 3/10 [00:03<00:06,  1.04it/s]
Epoch 38:  50%|█████     | 5/10 [00:04<00:04,  1.11it/s]
Epoch 38:  70%|███████   | 7/10 [00:06<00:02,  1.22it/s]
Epoch 38:  80%|████████  | 8/10 [00:06<00:01,  1.50it/s]
Epoch 38:  90%|█████████ | 9/10 [00:07<00:00,  1.38it/s]
Epoch 38: 100%|██████████| 10/10 [00:07<00:00,  1.75it/s]
                                                         
  Epoch 38: loss=0.4873 val_MRR=0.6870 val_R@10=0.8985 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 39:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 39:  10%|█         | 1/10 [00:02<00:18,  2.10s/it]
Epoch 39:  30%|███       | 3/10 [00:03<00:07,  1.07s/it]
Epoch 39:  50%|█████     | 5/10 [00:05<00:04,  1.08it/s]
Epoch 39:  60%|██████    | 6/10 [00:05<00:03,  1.32it/s]
Epoch 39:  70%|███████   | 7/10 [00:06<00:02,  1.11it/s]
Epoch 39:  90%|█████████ | 9/10 [00:07<00:00,  1.46it/s]
                                                        
  Epoch 39: loss=0.4725 val_MRR=0.6898 val_R@10=0.8946 val_Hit=96.6% alpha=0.9 (learned=0.996)

Epoch 40:   0%|          | 0/10 [00:00<?, ?it/s]
Epoch 40:  10%|█         | 1/10 [00:01<00:16,  1.80s/it]
Epoch 40:  30%|███       | 3/10 [00:03<00:07,  1.09s/it]
Epoch 40:  50%|█████     | 5/10 [00:04<00:04,  1.19it/s]
Epoch 40:  70%|███████   | 7/10 [00:05<00:02,  1.34it/s]
Epoch 40:  90%|█████████ | 9/10 [00:06<00:00,  1.65it/s]
                                                        
  Epoch 40: loss=0.4417 val_MRR=0.6873 val_R@10=0.8867 val_Hit=96.6% alpha=0.9 (learned=0.996)
  Early stopping at epoch 40 (no val improvement for 10 epochs)

============================================================
  Post-training final evaluation (on TEST)
============================================================
/workspace/ta-statute-law-retrieval/src/paragnn/trainer.py:280: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  model.load_state_dict(torch.load(f"{output_dir}/best_model.pt", map_location="cpu"))

  Alpha sweep on VAL (original):
  Alpha    R@10       MRR@10     Hit       
  --------------------------------------
  0.0      0.0884     0.0554     14.9%      <-
  0.1      0.1247     0.0720     22.3%      <-
  0.2      0.1708     0.1103     31.1%      <-
  0.3      0.2018     0.1502     37.2%      <-
  0.4      0.2148     0.2126     38.5%      <-
  0.5      0.2856     0.2895     48.6%      <-
  0.6      0.5234     0.4455     71.6%      <-
  0.7      0.8005     0.6110     91.2%      <-
  0.8      0.8838     0.6972     95.9%      <-
  0.9      0.9038     0.7136     97.3%      <-
  1.0      0.8625     0.6786     95.3%     

  Alpha sweep on VAL (debiased):
  Alpha    R@10       MRR@10     Hit       
  --------------------------------------
  0.0      0.0884     0.0554     14.9%      <-
  0.1      0.0988     0.0619     17.6%      <-
  0.2      0.1190     0.0677     21.6%      <-
  0.3      0.1382     0.0851     25.0%      <-
  0.4      0.1596     0.1076     28.4%      <-
  0.5      0.1731     0.1276     31.8%      <-
  0.6      0.1838     0.1583     33.8%      <-
  0.7      0.2148     0.2105     39.2%      <-
  0.8      0.3089     0.3066     49.3%      <-
  0.9      0.4838     0.4124     63.5%      <-
  1.0      0.5568     0.4447     70.3%      <-

  Grid Search (alpha=0.9, original, from val):
    Recall@10: 0.5864  MRR@10: 0.4448  Hit: 68.2%

Training complete. Test Recall@10: 0.5864
  Predictions saved: outputs/predictions/structgnn_nognn_statusaware_structdense_kuhperdata-exp.jsonl (296 queries, top-100)

  Final best Recall@10: 0.5864